Google could surpass Nvidia in AI accelerators by 2028

What happened?

According to a note from Fubon Research published by analyst Sean on X, Google plans to manufacture between 12 and 15 million units of its ninth-generation AI accelerator (TPU v9) by 2028. This figure would surpass the 12.4 million AI GPUs for data centers that Nvidia is estimated to sell in the same year, according to estimates from the same report. The report notes that the TPU v9 will use four compute chiplets, doubling capacity consumption compared to 2027. Tom's Hardware, which reported the note, highlights that Google was one of the first hyperscalers to develop its own custom AI processors about a decade ago and has been steadily increasing its deployment.

Why is it important?

Nvidia currently dominates the AI accelerator market, with an estimated share of over 80% of units sold for data centers, according to Mercury Research data from 2023. If Fubon's projections hold, Google would not only catch up to Nvidia in volume but could surpass it. This would have profound implications: Google would reduce its dependence on external suppliers, accelerate the development of its own AI models (such as Gemini), and pressure Nvidia to innovate faster. Additionally, using Intel Foundry to manufacture these chips, as suggested by Fubon's note, would be a boost for Intel's foundry business, which is seeking to position itself as an alternative to TSMC. It is worth recalling that Intel has been heavily investing in its Intel 18A process and advanced packaging technologies like EMIB and Foveros, which would be key to integrating the four chiplets of the TPU v9.

Consequences for the market and players

• For Nvidia: Although it maintains an advantage in performance and ecosystem (CUDA), losing the volume lead could erode its pricing power and market share in the long term. Nvidia already faces growing competition from AMD with its Instinct MI300 and from hyperscalers like Amazon (Trainium) and Microsoft (Maia). If Google manages to produce more accelerators than Nvidia, it could force Jensen Huang to accelerate product cycles or reduce prices.
• For Google: Vertically integrating accelerator production gives it control over its supply chain and optimizes costs, but the technical challenge of integrating four chiplets is enormous. Google already uses TPU v5p and v5e in its data centers, and the transition to v9 would represent a significant generational leap. Additionally, Google could offer TPU v9 capacity as a service on Google Cloud, directly competing with Nvidia's GPU-based instances.
• For Intel Foundry: Securing a client like Google for a high-volume product would validate its advanced packaging technology (such as EMIB and Foveros) and improve its credibility against TSMC. Intel has been struggling to attract external customers; a deal with Google would be a milestone comparable to TSMC's with Apple in 2014 for the A8 chips. However, Intel's ability to meet the required volumes (12-15 million units) remains to be proven, especially given that its foundry has not yet produced large-scale AI chips.
• For TSMC: It could lose some of Google's demand, although it would remain the primary manufacturer of AI chips for Nvidia, AMD, and others. TSMC currently produces Google's TPUs (such as the TPU v5p on its N5 process), but if Google migrates to Intel, TSMC could see reduced revenue from one of its largest customers. However, overall demand for AI chips continues to grow, which could offset the loss.

Historical and technological context

Google has been developing its own TPUs for a decade, starting with the first generation in 2015, initially designed for inference. The TPU v2 (2017) added training capability, and since then Google has released new generations every two to three years: v3 (2018), v4 (2021), v5e (2022), and v5p (2023). The decision to use four chiplets in the v9 reflects an industry trend toward modular designs to scale performance, similar to what AMD does with its chiplets in CPUs and GPUs (e.g., Ryzen and EPYC). It also echoes Intel's strategy with its Xe GPUs, which use tiles. Using Intel Foundry, if confirmed, would be a milestone: for the first time, a hyperscaler would trust Intel for a high-volume AI chip, breaking the near-total dependence on TSMC. Historically, Google has manufactured its TPUs with TSMC, but the need to diversify the supply chain and potential cost advantages could tip the scales.

“If Google manages to manufacture 15 million TPU v9s in 2028, we would be witnessing a paradigm shift: the largest buyer of AI accelerators would also become the largest manufacturer, surpassing the traditional supplier Nvidia.” — Fubon Research analyst (cited by Sean).

What readers should know

Fubon's projections are speculative and have not been confirmed by Google or Nvidia. The performance of the TPU v9 against Nvidia's Rubin architecture (expected in 2026) is unknown. Additionally, Intel Foundry's ability to meet the required volumes remains to be proven; Intel has had yield issues in the past with its advanced processes. However, the mere possibility that Google could surpass Nvidia in accelerator volume reflects the growing importance of custom chips in the cloud and the need to diversify manufacturing. It also highlights the intensifying race for AI supremacy, where hyperscalers are investing billions in custom silicon. For example, Amazon announced its Trainium2 chip in 2023, and Microsoft unveiled Maia 100 in 2023. Google, with its TPU experience, appears to have a maturity advantage.

Conclusion

The Fubon Research report paints a scenario where Google could challenge Nvidia's dominance in AI accelerators by 2028, even using Intel Foundry as a manufacturing partner. While many unknowns remain (actual volumes, performance, Intel's capacity), the trend is clear: hyperscalers are betting big on custom silicon, and Google seems determined to lead that race. The final decision will depend on technical execution and market evolution, but the mere fact that a production of 15 million units is being considered shows the magnitude of Google's ambition. The coming years will be crucial in determining whether Nvidia maintains its throne or Google manages to dethrone it with its own technology.